Unified Data Model

One deterministic understanding of the radiology universe.

SironaLex runs DICOM pixels, HL7 records, voice notes, and practice operations through semantic classification to build one unified model of every study and workflow. That single source of truth powers generalized AI, clinical automations, and complete practice visibility.

See Sirona in Action

How Sirona is Different

SironaLex: a custom radiology ontology, ML-assisted.

SironaLex is not a database schema. It's a custom OWL knowledge graph built for radiology, versioned in Git, and continuously improved by ML. The ontology captures every meaningful concept — anatomic regions (RadLex), clinical conditions (SNOMED CT), billing codes (ICD-10), modalities, protocols, severity — and the relationships between them. ML-trained NER models extract semantic triples from voice, dictation, and clinical text, and improve continuously as more studies are processed.

The unified data model ecosystem

From ingest to orchestration: every component flows data through the semantic ontology.

SironaLex Ontology

Custom OWL knowledge graph. RadLex, SNOMED CT, ICD-10. Versioned in Git. Continuously improved by NER models.

Classification Engine

FastAPI semantic reasoner over a SPARQL graph database. Real-time classification, confidence scoring, hierarchical filtering.

Data Ingestion Pipelines

DICOM standardization into versioned PostgreSQL. HL7 for EHR connectivity. Voice transcription and NER. Practice operational data.

Semantic Search & Filtering

Query by clinical meaning, not series description. 'All chest CTs with findings', 'all abdomen studies under 5mm slice thickness' — with confidence scoring.

Practice Intelligence

Admin dashboard with complete visibility into every study, workflow, and metric. Real-time analytics powered by the unified model. No silos.

AI Routing (Borvo)

Classification engine determines study type, protocol, and context. Routes to the right algorithms. Coordinates execution. Returns unified results.

Built for how radiologists actually work

Every layer — ingest, AI, analytics — flows through the same semantic model.

Data Pipeline

From ingestion to unified understanding

A study arrives: DICOM images, HL7 admission data, a voice note from the referring physician. SironaLex standardizes, extracts entities, and classifies the study against the ontology with confidence scores. The result is no longer files and text — it's a semantically understood entity with relationships to priors, similar studies, and clinical context.

DICOM standardization and metadata extraction

NER on voice, dictation, and clinical notes

Hierarchical classification against SironaLex

Confidence scoring and versioned reasoning trail

Graph relationships to priors and clinical context

AI & Clinical Workflow

Generalized intelligence built on unified data

The hanging protocol, AI stack, and clinical context all come from the same semantic model. Protocols key to semantic study type, not series description. AI algorithms route by classification and generalize across every practice. Clinical automations trigger on semantic conditions with confidence thresholds.

Hanging protocols keyed to semantic study type

AI routed by semantic classification, not hard-coded rules

Unified context for agents — images, reports, history in parallel

Automations trigger on semantic conditions

Generalization across every practice on the platform

Practice Operations

Complete visibility powered by semantic data

The admin dashboard doesn't query five systems — it queries the unified model. Volumes, complexities, SLA trends, turnaround times, and radiologist performance are all rooted in semantically unified data. The practice goes from flying blind to complete operational visibility in real time.

Real-time query across all modalities and findings

Volume, complexity, and performance by semantic type

SLA monitoring and bottleneck detection

Turnaround analytics broken down by study semantics

Trend analysis, forecasting, anomaly detection

Practice Intelligence

Practice intelligence on top of SironaLex

Because every study is semantically classified, the admin dashboard, real-time operational metrics, and AI routing all surface through the same semantic layer. Leaders query practice state directly — no pipelines to build, no reports to reconcile, no stale data.

Dashboards fed directly by the semantic layer

Real-time operational metrics across every workflow

AI routing decisions visible alongside clinical data

One semantic query, every answer

The unified data model advantage

ML

NER models improving with every study processed

1

ontology shared across every practice — AI generalizes

2018

unified data as the design principle since day one

Complete

operational visibility across every study and workflow

The impact of semantic understanding

Why Agentic AI Requires RadOS

Watch as Dr. Mark Longo demonstrates the power of a new paradigm in radiology AI. Welcome to the age of agentic, embedded, multimodal, real-time, clinically aware AI assistants.

Dr. Mark Longo

Chief Technology Innovation Officer, Sirona

Launch Day Excerpt: How Sirona is Built Different

Sirona's RadOS platform understands how a radiology practice really works. It starts by unifying all the data and tools needed for physicians to read seamlessly from anywhere at any time. The unique architecture and AI capabilities can automate many of the clicks, drags, scrolls, and “scratch thats” slowing your radiologists down.

See the platform in action

FAQs

What is SironaLex?

SironaLex is Sirona's custom OWL ontology for radiology — a knowledge graph capturing anatomic concepts (RadLex), clinical conditions (SNOMED CT), billing codes (ICD-10), and the relationships between them. It's versioned in Git and continuously improved by ML models trained on annotated reports. Every study gets classified against SironaLex to create a semantically unified understanding.

How does machine learning improve the model?

NER models trained on annotated radiology reports extract semantic triples from voice, dictation, and clinical text. Those extractions feed back into the ontology, improving classification coverage and accuracy. Every study you process makes SironaLex better.

Why can't legacy PACS vendors build this?

A unified data model requires unified infrastructure from the ground up. Legacy systems fragment data across on-prem PACS, separate RIS, disconnected dictation, and third-party AI. Retrofitting semantic classification onto that is cosmetic, not architectural. Sirona was built cloud-native with unified data as the design principle in 2018 — a foundational choice that created a widening moat.

How do hanging protocols work with the unified model?

Instead of a protocol per series description, you build one protocol per semantic study type — 'all chest CTs with standard inspiration protocol and multi-planar reconstruction'. The system applies it to every matching study regardless of how it was labeled at ingest. Hundreds of variations collapse to a handful, and protocols travel cleanly across practices.

How does the unified model enable AI generalization?

When AI is trained on studies unified by SironaLex, it generalizes across every practice on the platform. A lung nodule detector trained on 3,000 semantic chest CTs works everywhere. Without semantic unification, that same model would need retraining per practice. One training, universal application.

What visibility does the admin dashboard get from the unified model?

The dashboard queries the unified semantic model directly — not fragmented PACS, RIS, and AI systems. It sees volume by modality and finding type, complexity distribution, SLA trends, turnaround patterns, radiologist performance by semantic case type, and anomalies. All in real time.